Predicting Corrosion Rates of Carbon Steel in Real-World Aquatic Environments Using Machine Learning

Corrosion in aquatic environments cause significant economic losses and structural degradation. This study models the corrosion rate of S235 carbon steel using machine learning (ML) under real-world aquatic conditions. A field dataset from 46 locations along the Ghent–Terneuzen canal was used, encompassing exposure and environmental parameters such as temperature, pH, total dissolved oxygen (HDO%), chlorophyll concentration, oxidation–reduction potential (ORP), total dissolved solids (TDS), chloride concentration, specific conductivity, depth, and salinity. Six ML algorithms, including Light Gradient Boosting Machine (LightGBM), Gradient Boosting Regressor (GBR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Neural Network (NN), and Categorical Boosting (CatBoost) were benchmarked before and after feature selection. This work demonstrates that environmental feature selection provides substantially greater predictive improvement than model architecture choice: feature selection enhanced all algorithms from poor (R² ≤ 0.14) to strong performance (R² = 0.70–0.80), reduced inter-model variation by 64%, and decreased prediction error by 48% (RMSE) and 74% (MSE). LightGBM achieved the best performance (MSE = 0.003, R² = 0.80). Unexpectedly feature importance analysis identified, salinity, and depth traditionally considered critical factors showed minimal predictive influence, while exposure duration, pH, HDO%, temperature, chlorophyll concentration and ORP dominated corrosion behaviour. These findings emphasize the critical role of environmental parameters and feature selection over model complexity, supporting more efficient corrosion monitoring and management in marine and aquatic environments.